SShogo Learn
← Library

Guide · 2026-10-24

Six ways to run an agent in 2026 — the one-prompt build

What you'll have: the exact prompt that built the decision table in the reel, the steps to build it yourself in one message, and what to check before you trust what comes back.

What's real here: the prompt and the page. I sent the prompt below to a fresh Shogo project as one message, and the page in the video is what came back, recorded from the preview at phone width. One correction: after the build we fixed the two closing lines on the page itself, because the count in the prompt's verdict was wrong (more on that below). It isn't published anywhere. The table is our opinion of the six approaches, written by us line by line. The agent did the layout, not the opinions.

Who it's for

You've tried one or two ways of running coding agents and you're deciding what to commit to. Or you just want a clean example of a one-prompt build where you write every word and the agent does the design. Either way, you don't need to set anything up beyond a new Shogo project.

The steps

  1. Open Shogo and start a new project (the default web app is fine).
  2. Paste the prompt below as one message. Don't split it up and don't add a follow-up. The prompt already says what "finished" means.
  3. Wait for the reply. It comes back with a preview link. Mine took about seven minutes from sending to reply.
  4. Open the preview at phone width and at desktop width and go through the checks at the bottom of this guide.
  5. Optional: record it. I recorded mine at phone width (432×768 at 2.5×, so 1080×1920) for the reel. The page is static, so the scroll you see is a scripted pan down the real page, not an animation.

The prompt

Word for word what I sent, including the last sentence asking for the preview link. The same text is in shogo-guide.md next to this file.

Build a single-page decision table called "Six ways to run an agent in 2026". Mobile-first: the table becomes one stacked card per row on a 390px phone, and a real grid on a 1440px desktop. Dark UI, background #0B0B0F, orange #FB8C00 accents, Inter for text and JetBrains Mono for the column labels. No images needed.

Six rows, in this order, and four columns: who writes the code · who reviews it · who merges it · what it costs to leave. Use these exact cells, do not invent others:

1. Hosted agent — writes the code · nobody, until something breaks · you, after the fact · your prompts and your integrations
2. Local CLI agent — writes the code, in your shell · you, in review · you · nothing, it is your files
3. No-code workflow — the workflow does · nobody · you, by hand · every step you did not document
4. MCP server — the tool, when it is called · you · you · the client it was written for
5. Agent framework — your agent, once you wired it · whoever you built the reviewer for · you · the glue code
6. Agent workspace — a dev agent, on a card · a QA agent, before you · you · the board, every card stays

Give each row a small ordinal chip (01–06). Make the "who merges it" column the strongest visually: in five of the six rows that cell reads "you" and should read as the same answer repeating, with row 06 the only one where review happens before the merge instead of after it.

End the page with a verdict line in large type: "Five of the six end with you opening the pull request." and under it, small: "The merge is the part nobody automates away." Footnote: "Same six, from our own workspace."

It has to look finished on a 390px phone and a 1440px desktop, with no console errors and no placeholder text. When you are done, reply with the preview URL and the time the build took.

How it was built

One prompt, one build. I sent it as a single message to a fresh project on Oct 24, 2026. The agent built the page, checked it at four widths in a headless browser (390, 768, 1024 and 1440), killed the one 404 it found (a missing favicon), and replied with the preview link. I didn't send a second message.

One number in that reply is misleading. It answered "build time: 335 ms". That's the Vite build, not the seven minutes. When an agent reports a duration, check which clock it read.

I wrote the words; the agent did the layout. Every cell is in the prompt, with "use these exact cells, do not invent others". A comparison table is exactly where an agent will happily fill in plausible-sounding claims, so it gets no room to.

Why the merge column is the loud one. Read across any row and the first three columns change. The merge column barely does: it's you, every time. The real difference is whether anyone reads the diff before you merge, and when. In row 06 a QA agent reviews the work and opens the pull request before it reaches you. In rows 01 and 03 ("nobody, until something breaks" and "nobody"), nobody reviews it at all.

The verdict, as shipped. The prompt asks for "Five of the six end with you opening the pull request." That count was wrong: every one of the six "who merges it" cells says "you". So after the build we changed the page's closing lines to two lines you can check against the table:

  • verdict, in large type: "All six end with you merging the pull request."
  • under it, small: "In two of them, nobody read the diff first." (rows 01 and 03)

The footnote, "Same six, from our own workspace.", stayed the same. The reel was re-recorded from the corrected page. The prompt in this guide is still word for word what I sent, so if you run it as-is, you'll get the old "Five of the six" line and should fix it the same way.

What to check when it's built

  • Phone (390px): the table turns into six stacked cards, one per row, each with its 01–06 chip next to the name and all four answers labelled inside the card.
  • Desktop (1024px and up): it's a real grid, with the approach name plus four columns and mono column headers. Mine switched from cards to grid at 1024 and stayed a grid at 1440.
  • No horizontal overflow at any width. Nothing should scroll sideways on a phone. Drag the window from 390 to 1440 and watch for a sideways scrollbar.
  • The merge column reads as one repeated answer. "you", over and over, in the strongest style on the page. Row 06 should be the only row where review is visibly tied to the merge (on mine, the review cell joins the merge band), because it's the only one where review happens before the merge.
  • The verdict, sub line and footnote are there, word for word, and the console has no errors. Then correct the verdict's count, as described above: the prompt's "Five of the six" doesn't match its own table.

One thing to decide before you run it. The prompt asks for two things that don't quite agree: "use these exact cells" lists rows 01 and 03 as "you, after the fact" and "you, by hand", but it also asks the merge column to read as the same answer repeating. My build chose the repetition and showed plain "you" in all six rows, as the design note asked. That's how the shipped page stays, and it's also why the verdict's count had to change from "five" to "all six". If you want those two qualifiers on the page, take "should read as the same answer repeating" out of the prompt, or check the merge cells for rows 01 and 03 after it builds.

The honest caveat

"All six merge through you, two with no reviewer" is true of these six. Add a seventh row — a CI bot that merges on green, or any repo with auto-merge armed — and the merge stops being you. Row 06's you is a choice too: our own team leaves merges manual by default, so QA opens the pull request and I merge it. The table is an argument about where the last button sits, not a law of nature. If you build your own, say which six you picked, because that's the difference between a decision table and a survey.

How to use it

Open Shogo, start a new project, and paste the prompt (or shogo-guide.md) as one message. That's the whole setup. To make your own version, keep the structure and the "use these exact cells" line, and swap in your own rows.

Next step

Row 06 is how I run my own agents. The TEAM guide shows that setup: a planner, three developers and QA, and you merging at the end.

Set it up in Shogo

Open a Shogo workspace and paste this into the chat. Answer its questions; the agent builds the rest.

Prompt
Build a single-page decision table called "Six ways to run an agent in 2026". Mobile-first: the table becomes one stacked card per row on a 390px phone, and a real grid on a 1440px desktop. Dark UI, background #0B0B0F, orange #FB8C00 accents, Inter for text and JetBrains Mono for the column labels. No images needed.

Six rows, in this order, and four columns: who writes the code · who reviews it · who merges it · what it costs to leave. Use these exact cells, do not invent others:

1. Hosted agent — writes the code · nobody, until something breaks · you, after the fact · your prompts and your integrations
2. Local CLI agent — writes the code, in your shell · you, in review · you · nothing, it is your files
3. No-code workflow — the workflow does · nobody · you, by hand · every step you did not document
4. MCP server — the tool, when it is called · you · you · the client it was written for
5. Agent framework — your agent, once you wired it · whoever you built the reviewer for · you · the glue code
6. Agent workspace — a dev agent, on a card · a QA agent, before you · you · the board, every card stays

Give each row a small ordinal chip (01–06). Make the "who merges it" column the strongest visually: in five of the six rows that cell reads "you" and should read as the same answer repeating, with row 06 the only one where review happens before the merge instead of after it.

End the page with a verdict line in large type: "Five of the six end with you opening the pull request." and under it, small: "The merge is the part nobody automates away." Footnote: "Same six, from our own workspace."

It has to look finished on a 390px phone and a 1440px desktop, with no console errors and no placeholder text. When you are done, reply with the preview URL and the time the build took.
Want it as a guided course with a coach?Agent Team Playbook